VFCAnet: Voxel-wise Fusion and Channel-wise Attention Network for 3D Semantic Occupancy Prediction in Autonomous Driving

Engineering notes

Extensive experiments on the OpenOccupancy benchmark show that VFCAnet improves dynamic object recognition by 28.93% and boosts over-all mean Intersection-over-Union (mIoU) by 3.48% compared to strong baselines.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Accurate perception of 3D environments is crucial for ensuring safe autonomous driving. While traditional 3D object detection methods primarily focus on object localization, they often overlook critical background and environmental context. To address this limitation, we propose VFCAnet (Voxel-wise Fusion and Channel-wise Attention Network), a novel multimodal fusion framework for 3D semantic occupancy prediction. VFCAnet fuses camera and LiDAR features within the voxel space via a lightweight channel-wise attention mechanism, enabling fine-grained environmental understanding without the computational overhead of traditional convolutional operations. Extensive experiments on the OpenOccupancy benchmark show that VFCAnet improves dynamic object recognition by 28.93% and boosts over-all mean Intersection-over-Union (mIoU) by 3.48% compared to strong baselines. These results demonstrate the effectiveness of voxel-level attention-based fusion in advancing multimodal 3D perception for autonomous driving.

Links and sources

• Official / arXiv page

Comments